The question is not whether the AI boom ends. The question is which of four distinct paths it takes — and which signals distinguish one from another before the positioning damage is done. In The AI Boom: What Could Come Next1, Meketa Investment Group researchers Alison Adams, PhD, and Frank Benham, CFA, CAIA, argue that most current commentary has already chosen an ending — a leverage-driven bust — and that investors relying on that single narrative are watching the wrong indicators for the wrong event.
The Four End States
Adams and Benham separate the possible conclusions of the AI boom into four scenarios:
1) a financial break, in which the financing structure fails before assets generate sufficient returns;
2) a rotation, in which capital expenditure proves justified in aggregate but value migrates away from companies currently priced to capture it;
3) a transition, in which a break clears the path to a longer phase of falling costs and broadening adoption; and,
4) a demand disappointment, in which usage simply fails to grow into committed capacity without any accompanying credit event.
The authors are explicit that these are not competing forecasts: "These outcomes are not mutually exclusive, and they may occur one after another." A demand disappointment could trigger a financial break. A financial break could open into a transition. The analytical task is not to pick one — it is to know what evidence to watch for each.
The Circular Structure Nobody Is Fully Accounting For
The Minsky framework — in which borrowers move from servicing debt with operating cash flow, to rolling over principal, to issuing new debt simply to service existing obligations — is the financial break lens. Against roughly $1.1 trillion in AI-related capital expenditure by the four largest hyperscalers, some estimates now put the total data center buildout at $3 trillion, with as much as $800 billion in additional private financing potentially deployed over the next two years.
What makes the picture harder to read than a standard Minsky analysis is the circular financing structure running through the AI ecosystem. Adams and Benham describe it precisely: Nvidia invests equity in CoreWeave, which uses those proceeds to purchase Nvidia chips, with Nvidia then recognizing that purchase as revenue — while separately guaranteeing to buy any capacity CoreWeave cannot sell. "The same dollar can be recognized as revenue at more than one point in that circuit," Adams and Benham observe. "This implies that reported revenues may overstate actual demand."
Aggregate undisclosed obligations — SPV borrowings, uncommenced leases, and multi-year compute commitments structured as service agreements outside lease accounting altogether — have reached an estimated $1.65 trillion across the five largest hyperscalers. Until Anthropic and OpenAI go public, there is no balance sheet against which to measure their commitments at all.
When the Incumbents Are the Ones Most at Risk
The Schumpeterian framework asks a different question: not whether the financing holds, but who captures the gains. Adams and Benham make the portfolio implication direct: "The relevant point for investors is that Alphabet is spending because it believes that its current business model will not survive otherwise." This is not a cyclical capex decision. It is a survival judgment.
Chegg is the clearest example of the destruction side already in motion — paying subscribers fell from a peak of roughly 7.8 million to approximately 3.2 million by early 2025, revenue declined about 30% year over year, and the company cut close to half its workforce. No credit event, financing failure, or leverage was involved.
What a Transition Actually Looks Like
Carlota Perez's historical analysis of five technological revolutions provides the third framework. In each case, a financial collapse marked the boundary between installation and deployment — not the end of the technology's significance. The telecom buildout is the reference: the fiber stayed in the ground, bandwidth costs fell to levels that supported businesses that could not have existed at the prices the original carriers needed to earn. "What failed was the claims on the infrastructure, not the infrastructure itself," Adams and Benham note. A transition would appear as falling prices alongside rising volumes, capital expenditure decelerating toward the rate of general economic growth, and margin compression among current leaders while adoption broadens.
The Demand Condition Everything Else Depends On
The Jevons framework closes the analysis. As AI models improve and token costs fall, the question is whether demand expands faster than price falls. Anthropic's annualized revenue run rate is projected to reach $120 billion by the end of 2026, and margins on inference infrastructure have risen from 38% to over 70%. Demand is currently strong enough that Amazon recently raised the price of its EC2 capacity blocks by 20%. But the most aggressive capital expenditure plans rest on a proposition that remains untested: that AI agents will generate demand for other agents, with no obvious ceiling. "That proposition is plausible and, as yet, untested," Adams and Benham state. "It is also the single assumption on which the most aggressive capital expenditure plans depend."
Five Takeaways for Advisors and Investors
1. Monitor the gap between committed capacity and capacity in use. Utilization rates against committed capacity are more informative than token volumes alone. A widening gap is where a demand disappointment first becomes visible — before it becomes a headline.
2. Watch SPV spreads and off-balance-sheet obligations, not just reported debt. The five largest hyperscalers hold an estimated $1.65 trillion in aggregate obligations when SPV borrowings and uncommenced leases are included. Widening spreads on SPV-issued debt are an early financial break signal.
3. Capex deceleration is not bearish by itself — context determines the read. A slowdown in hyperscaler capital expenditure could mean the buildout has outrun demand. Read alongside falling token prices and rising usage, it signals that deployment is deepening and spending is normalizing toward general economic growth rates. The indicator is the same; the conclusion depends entirely on what accompanies it.
4. Concentration risk is present in all four scenarios. The companies currently priced as AI's primary beneficiaries occupy an unusually large share of public equity benchmarks. A rotation, a break, a transition, and a demand disappointment each reshuffle the industry differently — but all of them register in broad equity allocations. Concentration deserves examination as a risk separate from any view on whether AI proves transformative.
5. Check Anthropic's margin trajectory as a leading demand signal. Inference margins rising from 38% to over 70% while demand remains strong enough to push Amazon EC2 prices up 20% are not signs of a demand-constrained system. If those margins begin to compress as competition increases, the demand disappointment scenario is moving from theoretical to operational.
Footnote:
1 Adams, Alison, and Frank Benham. "The AI Boom: What Could Come Next." Meketa Investment Group, September 2026, https://meketa.com/leadership/ai-boom-what-could-come-next/.
